@elandau could you please take a look at this?

hi, bro, did u meet this situtation in issue #137 ?
i saw u use a final value 0.5 to degrade the gradient. how about using another smooth dynamic number which in (0,1) instead the final value ?
if u detect an drop, u cut down the gradient directly to 0.5. but if the dropped request is an accident, do u still want to cut it down?

@happyomg afaict, drop translates to back-pressure from the downstream service. In that case, using 0.5 as gradient is probably not a bad idea. 0.5 aligns with other limiters like AIMD in this repo. Remember that there is a smoothing factor to absorb any intermittent shocks.

@elandau do you have an estimate on when you would be able to review this? I am waiting on this since 8 days.

@elandau In case you are waiting for the travis failure fix, the failure seems to be a transient one.